Scholarlink

<p>Detecting malicious insider threats remains a major challenge due to the extreme class imbalance between normal and abnormal instances, which often leads to biased classifiers. To address this issue, we propose ACGAN, an adaptive loss–guided adversarial framework that integrates a generative adversarial network (GAN) with a convolutional neural network (CNN) classifier. The GAN’s generator produces realistic abnormal samples to balance the dataset, while the discriminator refines these samples through adversarial training. Moreover, we introduce two novel loss functions, range loss and pairwise comparison loss, to ensure that the generated data remains semantically meaningful and behaviorally consistent. The CNN classifier, trained on this enriched dataset, achieves significantly improved performance in detecting insider threats. Experimental results confirm that the proposed approach effectively mitigates data imbalance, thereby enhancing classification accuracy and model robustness.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ACGAN: Adaptive Loss–Guided GAN–CNN Framework for Semantically Coherent Synthetic Insider Threat Data Generation

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • Amel Slim,
  • Mehdi Malah

摘要

Detecting malicious insider threats remains a major challenge due to the extreme class imbalance between normal and abnormal instances, which often leads to biased classifiers. To address this issue, we propose ACGAN, an adaptive loss–guided adversarial framework that integrates a generative adversarial network (GAN) with a convolutional neural network (CNN) classifier. The GAN’s generator produces realistic abnormal samples to balance the dataset, while the discriminator refines these samples through adversarial training. Moreover, we introduce two novel loss functions, range loss and pairwise comparison loss, to ensure that the generated data remains semantically meaningful and behaviorally consistent. The CNN classifier, trained on this enriched dataset, achieves significantly improved performance in detecting insider threats. Experimental results confirm that the proposed approach effectively mitigates data imbalance, thereby enhancing classification accuracy and model robustness.